Auditing Qwen4-Max: 2.4T Parameters and a 60% Per-Task Claim
Alibaba's Qwen4-Max arrived in coverage dated 16 November 2026 with two numbers attached: 2.4 trillion parameters and roughly 60% lower cost per task. This audit separates what the release coverage actually states from what it leaves out — active parameters per token, unit pricing, the retry policy and the comparison baseline are all missing — and then builds a retry-adjusted cost-per-successful-task worksheet you can run against your own tasks this week. What's inside: - Why a per-task drop is a retry claim, not a rate cut: five channels can produce that 60%, and only one of them is a change to the price card. - MoE mechanics: 2.4 trillion total parameters is a memory story (about 4.8 TB of BF16 weights, 2.4 TB at FP8, 1.2 TB at 4-bit), not a per-token cost story. - Tokens per successful task, expected attempts under capped and uncapped retry loops, and the fixed per-task overhead most audits forget. - A metering wrapper that groups by task instead of by API call, so failed attempts stay in the numerator where they belong. No independent reproduction of the benchmark results was available at the time of writing, and no Qwen4-Max run backs these numbers: everything here is arithmetic over stated assumptions. Read the full write-up: https://hackyjs.com/posts/24t-parameters-60-cheaper-per-task-auditing-qwen4-max-after-retries